fix: detect nested MoE configuration

[Gwada]

Summary

• detect MoE expert counts stored under text_config
• preserve the existing top-level num_local_experts behavior
• add regression coverage for both configuration layouts

Problem

Some Hugging Face model configurations, including Qwen3.6 MoE checkpoints, store num_experts and num_experts_per_tok inside text_config. ModelProfiler only inspected top-level fields, so these models were classified as dense. As a result, their weight size was not measured correctly and --stream-experts was ignored.

Fix

Decode the nested expert fields in TextConfig and use them as fallbacks when the top-level values are absent. Top-level values keep priority, preserving existing behavior.

Validation

swift test --disable-sandbox --filter ModelProfilerTests

Executed 2 tests, with 0 failures.

The tests cover:

• nested text_config.num_experts and text_config.num_experts_per_tok
• the existing top-level num_local_experts format

Runtime validation with lmstudio-community/Qwen3.6-35B-A3B-MLX-8bit changed profiling from a false dense-model result to the expected MoE profile, allowing SSD expert streaming to activate.

[Gwada]

Closing as a duplicate of #114, which covers the same nested text_config case and also supports the other expert-count key spellings.